The questions your board will ask about AI
There are about six of them, they are predictable, and most companies can currently answer two. Here they are, with where each number comes from.
May 27, 2026 · The Camaze team
Board questions about AI have settled into a pattern. Two years ago they were about strategy: are we doing enough, are we behind. Now that the spend is material, they are about the number.
There are about six of them. They are predictable enough to prepare for, and most companies can currently answer two.
Here they are, with what a good answer looks like and where the number comes from.
1. What did it cost?
The easy one, and most companies still get it wrong, because the total they present is incomplete.
Three things routinely go uncounted. Model access billed through a cloud provider, which appears as a cloud line item rather than as AI. GPU compute running self-hosted models, which appears as ordinary infrastructure. And AI features inside SaaS contracts you already had, which appear as the SaaS contract.
A total that misses those is not a small amount wrong. It is frequently wrong by a large fraction, and the gap grows as self-hosting increases.
A good answer is one number, stated with confidence, that you know is complete because you can enumerate the sources behind it.
2. Where is it going?
The follow-up, and where the conversation gets useful.
Split by product and by department. What tends to emerge is concentration: two workflows accounting for the majority of the spend, with a long tail of small things that are not the problem.
That concentration is the most useful fact you can put in front of a board, because it converts an unbounded concern into two specific systems with owners. A board cannot act on the AI line. It can absolutely act on a conversation about two named workflows.
Include the two largest movements since last quarter, each explained in a sentence. Not a chart of everything. Two sentences.
3. What is it as a share of revenue?
For any company selling an AI-backed product, this is the question that matters most, and it is the one most likely to be asked by whoever on the board has seen a margin compress before.
AI is a variable cost that scales with usage, which makes it cost of goods sold, which means it hits gross margin directly. Reporting it in a general technology line hides that, and the hiding gets worse as the number grows.
The answer needs two parts: the current ratio, by segment, and the direction of travel. A ratio that is stable as you scale is a healthy business. A ratio that is deteriorating is a pricing problem that has not surfaced yet.
Be prepared for the follow-up, which is about the distribution rather than the average. Averages hide the accounts where the economics do not work.
4. What did we get for it?
The hardest one, and the one where most presentations become qualitative in a room where nothing else is.
The honest structure is per workflow, and it requires admitting that not every workflow has a clean denominator. Some do: cost per resolved ticket, cost per document processed, cost per enriched record, each set against what that unit is worth. Some do not, and for those the honest answer is a qualitative one plus the cost.
Do not construct a metric to fill the gap. A fabricated return figure does not survive the first follow-up question, and losing credibility on this question damages the answers to all the others.
A board is considerably more receptive to "these three workflows have measurable returns and here they are, these two do not have a clean measure and here is why we believe they are worth it" than to a single blended figure that nobody believes.
5. What does next year cost?
Give a range and state the assumptions. A single number implies precision you do not have, and this is a line where the honest uncertainty is wide.
Name what drives the width. Usually three things: rollout timing, because a feature shipping in Q1 rather than Q3 moves the annual number substantially; adoption rate within a rollout; and model pricing, which changes several times a year in both directions.
Show your prior forecast accuracy alongside it. Being able to say that intra-quarter projections have run within a few percent and annual projections within twenty is what makes the range credible instead of evasive. It also pre-empts the question about why you are giving a range at all.
If there are pending decisions with material cost implications, such as taking a feature to general availability, present those as scenarios rather than folding them into a single line. That turns the slide into a decision aid rather than a report.
6. How much of it could cost less?
This one is increasingly being asked directly, and how you answer it determines what happens next.
A large share of most AI bills can be delivered for less without reducing usage: a frontier model on work a smaller one handles identically, context that changes no output, caching and batching left switched off, agent loops retrying without a ceiling, workloads behind removed features, unused seats, discount tiers already earned.
If you present that figure alone, a board will reach for a cap. It is the rational response to a large unmanaged number.
If you present it alongside an itemized ledger with named findings, named owners, and an amount already realized and verified against actual spend, the same figure reads as a program under management. That is a completely different conversation, and it is the one that keeps the AI budget growing.
Report identified, in progress and realized separately. The difference between them is what tells a board whether the program is real.
What the slide should actually be
Four slides, and the discipline to stop there.
The number and the trend, split by product and department, with the two largest movements explained. The margin picture, by segment, with direction of travel. The forecast, with a range, stated assumptions and prior accuracy. The optimization program, with identified, in progress and realized.
Everything else is backup. The temptation with a new and uncomfortable line item is to over-explain it, and over-explaining reads as defensiveness. Four slides that answer the six questions crisply will do more for the AI budget than twelve slides of context.
The underlying point
The reason to be able to answer all six is not to survive the meeting. It is that a cost nobody can explain gets capped, and the cap is applied without the information needed to apply it well. It lands on the workflows that were working alongside the ones that were not.
Answering these six questions is what converts the AI line from something a board contains into something it invests in. That is the whole reason the exercise is worth the effort.